Experiments and comparisons

This page is derived from the canonical player-guide; its content has one source of truth.

Jev review and next experiments

After interpreting your setup, Jev checks the assembled species roles and food web against your original description. If its own proposal needs correction, LifePot attempts one bounded repair: Jev reselects the species and roles, then chooses feeding links against those fixed roles, and reviews the corrected graph again. Only an approved proposal can be seeded. Your answers are preserved; you do not have to rewrite them just to retry. If the correction remains unresolved, the review offers a specific model-selected area to clarify, or explains that no specific mismatch was identified. A failed correction stays blocked rather than silently becoming a fallback. You can explicitly choose the separate deterministic preset instead. The game does not silently add feeding links to make the story fit.

At runtime, the orchestrator and active scoped specialists use at most two provider requests per council observation. Specialist questions share a grouped inference request; responsibilities remain bounded, but these are not separate independent model calls. Token usage for the group is counted once. Existing separate-call replay records remain supported.

Reported adaptive abstentions distinguish rate limiting, unavailable service, invalid output, missing credentials, and unknown failures. An abstention applies no new policy; inherited behavior continues. A live setup does not guarantee every later decision is live.

In an adaptive world, Ask Jev pauses the current habitat and opens a compact proposal sheet. Drought, Nutrient bloom and Stability are editable request shortcuts: choosing one neither contacts Jev nor applies anything. Ask for a proposal, inspect the actual returned pressure, and explicitly accept it if you want to proceed. The returned pressure can differ from the shortcut or the request can be unsupported. Drought reduces regeneration; nutrient bloom increases it; stability reduces the hazard contribution rather than removing every hazard. None guarantees survival. Fixed starters remain inference-free unless you deliberately select adaptive policy before seeding.

An accepted change begins on the next engine generation, lasts six generations and restores the previous pressure. The world stays paused until you step or resume. Near the habitat, Recorded window compares actual population, mean resource and species measurements from acceptance through the observed window; the ending retains those observations before the raw histories. Missing exact boundary samples are not invented. If the world ends before the pressure window completes, its history says restoration was not observed rather than calling the change active. Comparisons distinguish starting-condition edits from recorded temporary player pressures and automatic Jev decisions; unchanged initial settings do not erase an intervention during observation. These are observations alongside a change, not proof it caused the differences; restoring a setting does not restore lost organisms. The optional private question and request prose are not added to portable exports.

Completed results include individual species histories. Edited follow-up trials put their changed-condition comparison before the individual run histories, retaining recorded conditions, aligned species trajectories and observed extinction timing. Conditions use readable labels; the rule-graph version is bookkeeping, not a second experimental condition. Starting or replaying a trial returns you to the board once, while later deliberate scrolling is preserved. Initial environmental hazard and graph pressure are separate conditions; changing one does not remove the other. Fresh adaptive decisions can differ even with the same initial seed, so comparisons do not establish causality.

Use Review fixed-rule baseline at the end of a run to review the original conditions and seed with runtime AI interventions disabled. This opens an editable review; it does not start another world or relabel the earlier adaptive run. Seed the separate fixed baseline, finish it, then use Edit this world to change one condition for a fixed-rule follow-up. Comparisons display actual seed and policy provenance plus counts and equality of accepted Jev decision records. Record differences include timing, observation and model metadata; they do not by themselves prove a different executed policy or its effect. Adaptive trials and recorded replays remain exploratory; different or unknown seeds never support a same-seed claim. Multiple changed conditions are exploratory, while unchanged fixed trials are repeat observations. Even a same-seed, single-condition fixed comparison describes this toy model, not biological proof.

Use “Ask Jev about this run” to request a model-selected turning point and supported next experiment. Jev selects among recorded evidence and bounded condition changes; displayed measurements come from the engine. Your optional private question is not sent. “Review suggested experiment” opens editable conditions without starting a run. If the service fails, no invented explanation is substituted. Evidence remains session-local; asking Jev sends numerical evidence and configuration to the server and model for that request.